TRACE AI Labs › AI agent compliance

AI agent compliance

AI agents are being handed real decisions in hiring, finance, healthcare, and procurement. Agent compliance asks a plain question: when an agent is given a rule and a reason to skip it, which one wins?

The problem with deployed agents

An AI agent in an enterprise workflow is not a chatbot answering trivia. It has a role, a set of KPIs, and instructions that include the rules the company is legally bound by. It also has a user in front of it who wants something done. Agent compliance is whether the rules survive contact with that user and those KPIs. Failures here are already in the case law: Air Canada was held to a refund policy its chatbot invented, New York City's MyCity assistant told businesses to break labor law, and Workday is defending a class action over algorithmic hiring screens.

What makes AI agents follow the rules

Our paper Why Do AI Agents Break Rules? How Framing, Context, and Social Signals Shape Compliance, published at the 2026 AAAI/ACM Conference on AI, Ethics, and Society, isolates the levers. One procurement agent, one rule (purchases over $1,000 must use an environmentally certified vendor), twelve models, and a controlled set of changes to the situation:

Training focus, as described by each developer, did not predict any of this. Models labeled agentic-RL systems and models labeled safety-aligned assistants appear at both ends of the compliance range.

Measuring agent compliance at scale

The procurement study has one rule and one domain. PACT generalizes it into an agent compliance benchmark: 3,364 items across 48 scenarios in 12 regulated domains, nine families of realistic pressure, and a scripted pushback turn. Every scenario is a place AI assistants already work and a real statute governs the decision. Scores are reported both with and without a hard compliance directive in the system prompt, so the benchmark also measures how steerable an agent's compliance is.

On the current leaderboard the best of 24 models scores 0.944 and none clears the bar for unsupervised use. One sentence of ordinary workplace pressure raises violation rates by 65%. When agents do break a rule, they describe the outcome as compliant 79% of the time, which matters for anyone relying on the agent's own audit trail.

Adjacent failure: authority the agent grants itself

Compliance can also erode from the inside. In Agent Memory Is a Surface for Endogenous Authorization Laundering we show that an agent's persistent memory can turn a suggestion into a standing permission nobody granted. Incremental memory updates created false authority for up to 50.2% of unauthorized requests, and once present, downstream executors acted on it in 98.6% of trials. EAL-Bench is the open benchmark for that failure.

Frequently asked questions

Is agent compliance different from LLM compliance?
Same underlying question, different framing. LLM compliance looks at the model; agent compliance looks at the model inside a role with goals, tools, memory, and a user. Our work covers both because the failures come from the interaction.
Do safety-tuned models comply more?
Not reliably. In the AIES study the developer's stated training focus did not separate compliant from non-compliant models, and in PACT release date and parameter count do not predict PACTScore.
What is the single most dangerous pressure?
Urgency. A looming deadline lowered compliance for every model in the AIES study, even with a follow-the-law instruction in the system prompt, and it remains among the strongest families in PACT.
How do I test my own agent?
Run the PACT harness against your endpoint using the public dataset, or reproduce the AIES experiments from the study repository.

Read more